Saund’s approach differentiates shape descriptions by building a vocabulary of symbolic shape descriptors using knowledge of the domain of the application. The classical approach builds shapes from a set of primitive building blocks applicable across a range of horizontal applications. Saund’s approach offers two general tools: a space-scale blackboard suitable for holding shape tokens with information about their relationships, and dimensionality reduction, which assists the interpretation of shape tokens as members of  families. 

The paper first reviews the drawbacks of the building block approach. It then describes how shapes decompose into fragments such as edges and corners with attributes like relative location, orientation, or scale. A family of shapes called a deformation class can be generated and mapped to a vocabulary of shape descriptors by varying parameters that control these attributes. The space-scale blackboard organizes shape tokens and serves as a working memory while rules are derived. Saund explains concepts using fish dorsal fins as examples.

Saund has made a significant contribution to shape recognition problems with a wide domain of applications varying from geographical information systems to CAD inference sketching. His argument that the building block approach is inferior because it has no knowledge of the domain is persuasive.

The paper introduces a large number of new concepts and terminology. Despite its length and complexity, the examples illustrate the concepts well. The paper is among the most readable I have seen.